Recursive Inference

Subagents receive fresh context windows, which can expand the effective context available to the system. Recursive inference is formulated as a dynamically generated execution tree with arbitrary branching up to a depth limit. Each recursive node uses the same policy instance to solve its assigned task. Recursive execution is presented as an inference-time scaling primitive for long-horizon agent tasks. Recursive rollouts can create related child tasks that form an implicit curriculum and provide intermediate credit-assignment signals.